Local machine without GPU acceleration

Do’s:

• Use numbas jit features in a fire and forget manner. Just decorate your functions with a numba.jit and see if this already does the trick.

• Or make use of the numpy api as often as possible to get nuerical focus and faster execution.

• Use the divide and conquer pattern to split up the workload and use the concurrent.futures interface to create a ProcessPoolExecutor executing it in parallel.

• Move forward to more specialized libraries implementing a numpy like interface. The changes to make here will be usually very subtle and can be done iteratively without eating to much time.

Dont’s:

• Do not use ThreadPoolExecutors due to the GIL.

• Do not mix numba and concurrent.futures based approaches, as usually numba will target all available cores already